Targeting the Undruggable: Deep Learning-Driven Design of Peptide Therapeutics in Cancer

The majority of disease-associated proteins are considered "undruggable" due to the absence of well-defined binding pockets, the presence of extended interaction surfaces, and intrinsic structural disorder, which collectively limit the effectiveness of conventional small molecules and biologics. Representative examples include KRAS, p53, and c-MYC. Peptide therapeutics, particularly macrocyclic peptides, occupy a unique chemical space capable of targeting such recalcitrant protein-protein interactions (PPIs) where small molecules often fail. However, traditional peptide discovery, which relies heavily on high-throughput screening, is labor-intensive and frequently yields candidates with suboptimal pharmacological properties. The integration of artificial intelligence has begun to transform peptide discovery from a largely empirical process into a rational and design-driven paradigm. Modern deep learning approaches, including diffusion-based generative models, enable the de novo design of peptide binders with high affinity and structural precision, even for disordered or previously intractable targets. In this perspective, we highlight key structural and biological challenges associated with undruggable proteins and consider how peptide-based modalities are beginning to overcome these longstanding barriers. We further explore how advances in artificial intelligence and computational modeling may reshape the rational design of next-generation peptide therapeutics and propose an integrated experimental-computational framework to facilitate the development of clinically actionable candidates.